Synthetic persona: what research validates and what it does not
A synthetic persona is a customer profile simulated by a language model, one you can question like a survey respondent. Academic research shows it reproduces some distributions of human answers accurately and fails on others. It serves to prepare and explore, never to replace a decision grounded in real customers.
- A model reproduces what people say publicly, not what they do when it comes time to buy.
- It excels at generating hypotheses and objections to test, never at validating them.
- Its biases come from its training data: populations over-represented, under-represented, absent.
- Used to settle a decision, it manufactures false confidence at zero cost.
On this page
Synthetic persona
A synthetic persona is a representation of a customer segment simulated by a language model, one you can ask questions of as you would a human respondent. It differs from the traditional persona, a fixed descriptive document, by its ability to produce new answers. It differs from a real customer in that it reproduces regularities of discourse rather than lived experience.
What research shows, and what it does not
The topic has produced a great deal of commercial enthusiasm and few careful readings. Serious academic work does exist.
Researchers have shown that a language model conditioned on sociodemographic characteristics can reproduce, with notable fidelity, some distributions of answers observed in real surveys. The term used in the literature is silicon sampling. Other work has simulated entire populations of agents to observe collective behaviour.
These results are real, but they call for three caveats that commercial promotion systematically leaves out.
The model reproduces discourse, not behaviour
It learns from what people have written, so from what they say about themselves. Yet the gap between what a person states and what they do at the moment of paying is precisely the subject of marketing.
Fidelity varies enormously by group
It is better for populations abundantly represented in the training data and degrades for the others. An industrial SMB in Saint-Jérôme is not an over-represented segment on the web.
The model never says it does not know
It produces a plausible answer in every case, with the same assurance. Nothing in the output distinguishes a faithful reproduction from a coherent invention.
The useful question is not whether it works. It is what you would do differently if the answer were false. If the answer is nothing, the use carries no risk. If it is you would change your offer, you need real customers.
What to know before relying on this
The appeal of this method is its cost: questioning a simulated profile costs almost nothing, while a series of customer interviews costs weeks. That is precisely what makes it dangerous. It produces confidence at zero cost, and a business decision made on that basis costs far more than the interviews it replaced.
- What does this profile rest on, data from our real customers or an imagined description?
- Which decisions do we intend to make from this, and which do we refuse to make this way?
- How many real customers have we interviewed this year, and when was the last time?
- Will what comes out of it be tested against real people before any spending is committed?
- Would five real interviews truly cost more than the decision we are about to make?
The useful answer treats each output as a hypothesis to verify and plans for the test against reality. A weak answer presents the results as a study, speaks of validated segments, or offers to replace customer interviews.
Deciding what can be settled by simulation and what demands real customers avoids a costly mistake. A 90-minute consultation settles it, with a written summary you can pass around your organization.
Where it is genuinely useful
Three uses hold up, and they share one feature: the error costs little there, because a verification always follows.
| Use | What it produces | Mandatory verification |
|---|---|---|
| Generate objections to anticipate | A broad list, some of them real | Test them against your sales conversations |
| Prepare an interview guide | Questions you would not have thought of | The interview with real customers |
| Explore message angles | Wordings to test | A real test in a campaign |
| Simulate a critical reading | The unclear zones of a page | A human usability test |
| Broaden a poorly understood segment | Hypotheses about needs | Discovery interviews |
The second use is by far the most profitable. Preparing a customer interview with the help of an AI produces a more complete guide in thirty minutes than a team builds in half a day. And the real interview itself remains irreplaceable.
Where it misleads
The risk is not that the AI gets it wrong. It is that it gets it wrong in a plausible, coherent and cost-free way, which makes verification unattractive.
| Use | Why it misleads |
|---|---|
| Validate a price | A model has never had to pay. It produces a stated preference, with no real trade-off. |
| Choose between two offers | Stated preferences and real choices diverge, an established finding in research |
| Estimate a market size | No basis: the AI does not know your local market |
| Replace customer interviews | You lose exactly what you were after, lived experience |
| Justify a decision already made | The model will confirm, because it follows the wording of the question |
The last case is the most frequent and the hardest to spot. A model follows the way a question is asked. A question that assumes the right answer gets the right answer, and the document it produces looks like a validation.
What stays in-house is access to real customers. Five interviews are enough to reveal the essential according to the Nielsen Norman Group, and no one on the outside can secure those meetings in your place: they run through your salespeople and your existing relationship. It is the resource this method claims to replace and does not. What gets delegated is the rest: building the profiles from your real data, framing unbiased questions, sorting the hypotheses and designing the verification protocol.
How to use it without going wrong
Here is the method we apply when an engagement justifies this kind of exploration.
Start from real data, not an imagined description
Feed the AI real verbatims: incoming requests, service exchanges, customer reviews, questions asked on calls. A persona built on your own data is worth infinitely more than a profile described from memory.
Ask open, unbiased questions
"What would make you hesitate" rather than "why is this offer appealing." The second wording guarantees a useless answer.
Treat each output as a hypothesis
An objection produced by an AI is not an observed objection. It becomes a line to verify with three real customers.
Always test against the human
Five real interviews are worth more than five hundred simulated answers. The Nielsen Norman Group has long documented that five people are enough to reveal most of an interface's problems.
A synthetic persona is an excellent generator of questions and a poor supplier of answers. Confusing the two is costly, and it does not show right away.
Falia analysis gridBefore using a simulated persona
Building a persona from real data is covered in buyer persona. The mechanics of purchase decisions, and the gap between what is stated and what is done, are detailed in the drivers of an online decision. The use of AI in execution is covered in what AI actually changed about content.
This point sits inside the framework described in who measures what, and who measures nothing.
The rigorous use of AI in execution is at the heart of the Strengthen your visibility in AI answers goal.
Already running a marketing team? See how we plug in as reinforcement on conversion rate optimization.
Frequently asked questions about synthetic personas
What is a synthetic persona?
It is a representation of a customer segment simulated by a language model, one you can ask questions of as you would a respondent. It differs from a traditional persona by its ability to produce new answers, and from a real customer in that it reproduces discourse rather than lived experience.
Can it replace customer interviews?
No. A model learns from what people write, so from what they state, whereas the gap between the stated and the done is precisely the subject of marketing. Five real interviews are worth more than five hundred simulated answers.
Does research validate this approach?
Partly. Academic work shows that a language model conditioned on sociodemographic characteristics reproduces some distributions of survey answers. Fidelity varies strongly across groups and degrades for populations poorly represented in the training data.
Can it be used to validate a price?
No. A model has never had to pay. It produces a stated preference with no real trade-off, whereas the divergence between stated preference and actual choice is an established finding in consumer behaviour research.
What is its best use?
Preparing a customer interview guide. It produces in thirty minutes a questionnaire more complete than a team builds in half a day, and the real interview that follows remains entirely irreplaceable.
How do you avoid sycophantic answers?
By asking open questions that do not steer. "What would make you hesitate" rather than "why is this offer appealing." A model follows the wording of the question, and a question that assumes its answer gets it.
- Lisa P. Argyle et al., Out of One, Many: Using Language Models to Simulate Human Samples, Political Analysis, Cambridge University Press, 2023.
- Joon Sung Park et al., Generative Agent Simulations of 1,000 People, Stanford University, arXiv, 2024.
- Jakob Nielsen, Nielsen Norman Group, Why You Only Need to Test with 5 Users, accessed July 2026.
Geneviève puts the strategy for your engagement into action. She leads all our web development projects: Shopify, WordPress and the new ways of building a site with AI. She manages our team of developers and translates your business needs into technical language. She runs your organic search (SEO), your visibility in AI answers (GEO) and your site's conversion rate optimization (CRO). Her work is at the heart of three goals: Attract customers with SEO and AI, Improve your site's conversion, and Strengthen your visibility in AI answers. With Gabriel, she also builds the landing pages for your advertising campaigns. She writes mainly about SEO, AI visibility and web design.
About Falia →
SEO
Paid advertising
Conversion
AI visibility
Web design